515 Artificial Intelligence and Whole Slide Imaging in Diagnostic and Prognostic Challenges of Melanoma; A Systematic Review

Introduction/Objective Artificial Intelligence (AI) and whole-slide imaging (WSI) have shown promising performance as a support tool in diagnosis, prognosis, and reducing interobserver variability. Methods/Case Report 21 studies using WSI, published between 2015 and 2025, were selected from PubMed and Google Scholar. Dermoscopic or non-cutaneous melanoma was excluded. Models used included CNNs such as ResNet, VGG16, and Inception, as well as transformer architectures and contrastive learning. Results Included studies were categorized into diagnostic prediction (n = 12), prognostic modeling (n = 7), and interobserver comparison (n = 2). The most commonly used models were CNNs, such as ResNet (Microsoft), VGG16 (Oxford), and Inception (Google), as well as transformer architectures (Google) and contrastive learning methods like SimCLR (Google). The number of WSIs ranged from 43 to 52,202. The most common model was ResNet or its variants. The highest diagnostic performance was achieved in a study with ResNet50 on 701 WSIs (AUROC = 0.971). The best prognostic result was obtained from a study with an AUROC of 0.905, and the strongest interobserver modeling achieved a Pearson r of 0.589 using the SimCLR model. Conclusion AI models, particularly CNN-based architectures, demonstrate strong diagnostic and prognostic capabilities in melanoma WSI analysis, offering performance equal to or surpassing expert pathologists.